@plur-ai/mcp

Give your AI agent persistent memory. One line in your MCP config — corrections, preferences, and conventions persist across sessions. No workflow changes, no cloud, no API costs for search.

Part of PLUR — where Haiku with memory outperforms Opus without it at 10x less cost.

Setup (30 seconds)

Claude Code

One command — sets up storage, MCP config, and hooks:

npx @plur-ai/mcp init

Restart Claude Code. Done. Your agent now has persistent memory with automatic injection.

Cursor

Add to .cursor/mcp.json:

{
  "mcpServers": {
    "plur": { "command": "npx", "args": ["-y", "@plur-ai/mcp"] }
  }
}

Windsurf / any MCP client

Same pattern — point it at npx -y @plur-ai/mcp.

That's it. Your agent now has memory. Use your tools as usual — corrections accumulate automatically.

What happens

You correct your agent  →  engram created       →  YAML on your disk
Next session starts     →  relevant ones injected →  agent remembers
You rate the result     →  engram strengthens    →  quality improves

Knowledge is stored as engrams — small assertions that strengthen with use and decay when irrelevant. Search is fully local (BM25 + embeddings), so memory recall costs nothing and works offline. Benchmark methodology →

Tools

Your agent gets these tools automatically:

| Tool | What it does | |------|-------------| | plur_session_start | Start a session — injects relevant engrams for your task | | plur_learn | Store a memory — correction, preference, convention, or decision | | plur_recall_hybrid | Best default — BM25 + embeddings merged via RRF. Zero cost. | | plur_recall | Keyword search (BM25 only, instant) | | plur_inject_hybrid | Load relevant memories for the current task | | plur_feedback | Rate a memory — trains relevance over time | | plur_forget | Retire a memory (history preserved) | | plur_session_end | End a session — captures summary and new learnings | | plur_capture | Record a session event | | plur_timeline | Query session history | | plur_ingest | Extract learnings from text | | plur_sync | Sync memory across machines via git | | plur_packs_install | Install a shareable memory pack | | plur_packs_list | List installed packs | | plur_status | System health |

Sync across machines

Your agent can sync memory to any git remote:

Agent: plur_sync({ remote: "[email protected]:you/plur-memory.git" })
→ "Initialized and pushed."

# On another machine, same remote:
Agent: plur_sync()
→ "Synced. Pulled 12 remote commits."

Works with GitHub, GitLab, Gitea, any git host. Your data, your repo.

Configuration

Custom storage path:

{
  "mcpServers": {
    "plur": {
      "command": "npx",
      "args": ["-y", "@plur-ai/mcp"],
      "env": { "PLUR_PATH": "/path/to/storage" }
    }
  }
}

Default: ~/.plur/. Everything is plain YAML — open it, read it, edit it.

Benchmark

| Metric | Score | |--------|-------| | LongMemEval overall | 86.7% | | A/B win rate vs no memory | 89% | | House rules accuracy | 100% |

Full methodology →

Related packages

| Package | For | |---------|-----| | @plur-ai/core | Engine — use directly in custom agent frameworks | | @plur-ai/claw | OpenClaw — automatic memory without MCP |

License

Apache-2.0 · GitHub · plur.ai